Who Is in Your Category? The Same Market Tracked Two Ways Gave Two Different AI Leaders

Most debate about Generative Engine Optimization (GEO) measurement is about sampling: how many prompts, how many days, which engines. Those are real problems. But there is a quieter variable that sits upstream of all of them and is almost never audited: which brands you decided to count. Mention share is a share of something, and the something is a list somebody typed in.

A public dataset happens to contain a clean natural experiment. It tracks two pairs of near-identical categories under separate pages: two for programming schools and two for English learning. Same engines, same collection window, same day of retrieval. The only thing that differs is the tracked brand set and, presumably, the query mix. The results diverge enough to change the answer to "who leads?" in one pair and to move every brand's number in the other.

The Data and Its Limits

The figures come from AIOPulse's public category pages, retrieved on 21 October 2026, all last updated on 20 October. They are produced by automated daily querying of ChatGPT, Gemini and Claude in Japanese at 06:00 JST. Each page covers an 8-day ranking period with 16 valid data points: four days each from ChatGPT and Gemini, which are retrieved every other day, and eight from Claude. Failed collection days are excluded from averages rather than counted as zero. The blended figure is an unweighted mean of the three engines, and shares are among the tracked brands only, so they sum to roughly 100% by construction.

One limit matters for everything below: the pages say they send "representative queries" but do not publish the prompt text. I can see the brand lists, which differ, but I cannot see whether the prompts differ too. Where I attribute a gap to the brand list, I will show the arithmetic; where a residual remains, I will say it is unexplained.

Pair One: Programming Schools, Six Brands Versus Seven

The first programming-school page tracks six brands. The second tracks seven: the same six plus CodeCamp.

Brand	Six-brand page	Seven-brand page	Seven-brand, CodeCamp removed and re-scaled
DMM WEBCAMP	40.2%	34.7%	43.6%
CodeCamp	not tracked	20.5%	—
Tech Academy	18.4%	17.6%	22.1%
Tech Camp	13.9%	8.1%	10.2%
Samurai Engineer	12.8%	7.5%	9.4%
RUNTEQ	10.1%	7.8%	9.8%
Potepan Camp	4.4%	3.7%	4.7%

The re-scaled column is my own calculation: take the seven-brand shares, drop CodeCamp's 20.5%, and divide each remaining brand by the 79.5% that is left. If the brand list were the only difference, that column would match the six-brand page. It gets close but not all the way. Two things are worth reading off it.

• Adding one brand cut everyone else's share, mechanically. CodeCamp takes 20.5 points from a fixed pool of 100. DMM WEBCAMP "lost" 5.5 points between pages without a single answer changing in its disfavour. A team that switched tracker configurations mid-year and added a competitor would see a drop of exactly this kind and could easily mistake it for a real decline.
• The denominator explains most of the pattern, but not all of it. Re-scaling lifts DMM WEBCAMP from 34.7% to 43.6%, which overshoots the six-brand page's 40.2% by 3.4 points. Samurai Engineer and RUNTEQ swap places (9.4% versus 9.8% re-scaled, but 12.8% versus 10.1% on the six-brand page). Whatever produces that residual, whether different prompts, different samples or different brand-name matching, it is as large as many of the movements people report in month-to-month dashboards.

The ChatGPT row shows the sensitivity most sharply. On the six-brand page ChatGPT gives DMM WEBCAMP 65.0%, and RUNTEQ and Potepan Camp both 0.0%. On the seven-brand page ChatGPT gives DMM WEBCAMP 44.7%, CodeCamp 33.4% and, again, zero to RUNTEQ, Potepan Camp and Samurai Engineer. The same engine over the same week is "dominated by one brand" on one page and "a two-horse race" on the other.

The 30-day trend panels disagree too, which is a warning about trend lines in general. Samurai Engineer is shown at −11.7 points on one page and +0.8 on the other; Tech Academy at −6.7 and +0.4. The end-point values on these panels are clean round numbers such as 8%, 12% and 24%, the signature of a single day with only a handful of mentions behind it. Trend endpoints built from one day of data are the weakest numbers on either page.

Pair Two: English Learning, Where the Leader Changes

The English conversation page is built around online lesson services and established schools. The broader English-learning page adds Duolingo and PROGRIT and drops Cambly and ECC.

Brand	Conversation page (blended)	Learning page (blended)
DMM Eikaiwa	30.2% (1st)	22.4% (2nd)
Native Camp	26.6% (2nd)	16.1% (3rd)
Rarejob	20.4% (3rd)	15.1% (4th)
Cambly	13.2% (4th)	not tracked
Studysapuri ENGLISH / Studysapuri	3.2%	11.8%
Berlitz	3.0%	2.8%
ECC	3.2%	not tracked
Duolingo	not tracked	23.5% (1st)
PROGRIT	not tracked	8.2%

Three observations, in descending order of how much they should worry you.

The leader depends on the list. A conversation-school marketer reading the first page concludes that DMM Eikaiwa owns the category with a 3.6-point lead. A language-app marketer reading the second concludes that Duolingo is ahead. Both are accurate descriptions of their own tracked set. Duolingo's lead on the second page is 1.1 points over DMM Eikaiwa, inside any reasonable noise band, so the honest reading is a contested top two. The point is that the first page cannot tell you Duolingo is a threat, because Duolingo is not on it. A tracker cannot report competition from a brand you did not list.

The same brand can mean different things on different pages. The service appears as "Studysapuri ENGLISH" on one page and plain "Studysapuri" on the other, at 3.2% versus 11.8%, a gap of nearly four times. I cannot tell from the public pages whether that is the product line, the umbrella brand or simply different prompts surfacing different parts of the company. It is exactly the kind of alias decision that a vendor makes silently and a buyer should ask about: when an answer says "Studysapuri", does that count toward the English product, the whole brand, or neither?

Engine splits widen when the list widens. On the learning page Duolingo is first on ChatGPT at 40.0% and fifth on Gemini at 9.8%, a 30-point spread. PROGRIT is first on Gemini at 24.6% and records zero on both ChatGPT and Claude. The ChatGPT column also reads 40, 20, 20, 20 across four brands and zero for the remaining three. Shares that round are what a few mentions look like; the page does not publish raw counts, so treat that column as a coarse indicator. The mechanics of this kind of divergence are covered in our earlier piece on per-engine splits.

Why This Is a Bigger Problem Than It Looks

Mention-share dashboards typically ask you, once, at setup, to list your competitors. That setup screen then determines every number you will report for a year. Three structural facts follow.

• Shares are closed-world. Total is fixed at 100, so any addition or removal re-prices every incumbent. The programming-school pair shows a seven-versus-six-brand swing of up to 5.5 points for the leader with nothing else changed.
• Industry boundaries are not the model's boundaries. A language-app and a live-tutoring marketplace sit in different industry codes but compete for the same buyer's question. If you build your list from your sales team's rivals, you will systematically omit the ones that arrive from the side.
• Not tracked does not mean not mentioned. The pages say this explicitly: untracked brands are shown as unmeasured, not absent. A leader at 40% may simply be a leader among the brands somebody remembered to list.

A Category-Definition Audit You Can Run This Week

1. Write down the brand list and who chose it. If nobody can say, you do not have a baseline, you have a default.
2. Add one adjacent competitor and re-scale. Pick a brand from a neighbouring category that a buyer might plausibly consider instead of you. Run the comparison and see how your share and rank change. If your lead disappears, your lead was a function of the list.
3. Re-scale before you compare periods. Whenever the tracked set changes, recompute historical shares over the common brand set before reporting any trend. The re-scaling formula used above is one line in a spreadsheet.
4. Test the aliases. For each brand, list every name a model might use (product line, parent company, abbreviation, English and local-language forms) and confirm how the tool matches them. Ask what happens to a mention of the parent brand.
5. Run a second prompt set on a subset. Take the same brand list and ask a differently-worded question. If the leader changes, the metric is measuring the question as much as the market.
6. Distrust one-day trend endpoints. Any trend panel built on a start-day and end-day pair is a two-sample comparison. Prefer windowed averages, as discussed in our piece on reporting windows.

What This Means for Tool Selection

The relevant question is no longer which tool has the nicest dashboard but how much control you have over the list and the prompts. The established Western platforms, including Profound, Peec AI and Otterly.ai, are built around you defining your own prompts and tracked competitors, which is the right default for this problem. The trade-off is that you carry the setup burden, and a poorly chosen list will give you a confident wrong answer in any of them.

AIOPulse, whose public pages supplied the data above, is a narrower product: ChatGPT, Gemini and Claude only. Its public category pages use a fixed brand list and unpublished representative queries, which is why this article could compare them but could not fully explain the differences. Its paid site tracking plan is listed at JPY 19,800 per site per month excluding tax with five keywords included, additional keywords at JPY 5,000 each per month, no initial fee and no free trial; the first month is charged at signup, and contracts run month to month with no partial-month refunds. Its FAQ says keywords can be added or removed at any time and that competitor setup is optional. It does not cover Perplexity, Copilot or AI Overviews citations. In Japan, Mieruca GEO is the better-known broader suite. Verify current prices and capabilities directly with each vendor before buying, because this category changes quickly.

FAQ

Which of the two pages is "right"?

Neither is wrong; they answer different questions. The right page is the one whose tracked set matches the set of brands your buyers actually compare. If you cannot tell, run both and report the range.

Should I track as many competitors as possible?

Not indefinitely. More brands dilute every share and make small differences harder to read, and the list gets harder to maintain. Aim for the brands that appear together in the answers you care about, plus a small rotating group of adjacent challengers.

Does this apply outside Japan?

The mechanism does. Closed-world shares, alias matching and prompt wording are properties of how any mention-share metric is built. The specific numbers here are Japanese-language and say nothing about your own market; use them to calibrate how much a brand list can move a result.

Takeaway

A mention share is only as meaningful as the list of brands underneath it. In the public data above, adding or swapping a handful of names moved a leader's share by up to 5.5 points, flipped the category leader in one case and changed who looks like the biggest threat. Before you debate sampling or engine coverage, audit who is on the list, how each brand's name is matched, and whether your closest adjacent competitor is missing entirely.

Disclosure: AIOPulse is developed by Marketing Cloud Lab, which also operates this site. All AIOPulse figures in this article were retrieved on 21 October 2026 from its public category pages (last updated 20 October 2026, 8-day ranking period), and its pricing was read from the published pricing section on the same date. The re-scaled column in the programming-school table is the author's calculation. Competitor coverage for Profound, Peec AI, Otterly.ai and Mieruca GEO reflects general public information as of October 2026 and was not independently audited.